Luigi Salemme

Developer, VFX Artist and AI enthusiast

[Luigi Salemme @0xLuigislm]

@macintogdev @ivanfioravanti Will look into it! Thanks for sharing

[Luigi Salemme @0xLuigislm]

@macintogdev @ivanfioravanti I already have fan control software and that’s how I see that setting them to 100% keeps the mac from throttling. I was wondering if there was another way that wouldn’t sound like a jet engine on my desk 😅

[Luigi Salemme @0xLuigislm]

@macintogdev @ivanfioravanti Any solution to this? Hope is just a temporary issue

[Luigi Salemme @0xLuigislm]

@ivanfioravanti I'm having a decent amount of issues. Stuttery bluetooth being the main one.

Do you notice your mac throttling harder on MacOS 27 by the way? Feels like it gets heat soaked faster and throttles much earlier than on MacOS 26. Only way to avoid that is to set fans to 100% speed

[Luigi Salemme @0xLuigislm]

@ivanfioravanti @ComfyUI Perhaps you might be interested. This setup would be plug and play, does not require per model adaptations (only optimizations to improve speed).

[Luigi Salemme @0xLuigislm]

I've been doing custom kernel work on ComfyUI, releasing on GitHub soon.
Results on my M5 Max: over 2x speed vs vanilla ComfyUI on most models, full ComfyUI functionality retained, still on MPS (no MLX backend swap).
Close to or matching DrawThings Speeds, faster than MLX ports.

[Luigi Salemme @0xLuigislm]

I like DrawThings, but ComfyUI's flexibility is still essential for a lot of my workflows, so I've been on a mission to get it running properly on Mac.
More models to test in the coming days. If you're on M4/M3 and want to help test once this drops, let me know, i'll need testers

[Luigi Salemme @0xLuigislm]

Benchmarks (M5 Max, 1MP unless noted):• ZiT — 8 steps: 7.5s
• Flux.2 Klein 9b: 4 steps ~12s edit / ~5s t2i
• Krea 2 Turbo: 8 steps ~14-15s
• SDXL: 12 steps ~3.5-4s
• LTX 2.3 (1024x1024, 121f): 8+3 steps ~100s
• Wan 2.2 (512x512, 81f): 2+2 steps: 120s

[Luigi Salemme @0xLuigislm]

@ivanfioravanti Will share results on my profile soon. Numbers are very promising and I think there might be room for improvement. The fun part is that it’s all with MPS, no MLX, so the whole comfy pipeline still works as is. All plug and play.